Context
Large inventory systems contain items with very different demand patterns, lead times, prices, and service requirements. Applying one safety-stock formula to every item is easy to operate but can be inaccurate: a conservative rule ties up working capital, while an aggressive rule creates shortages and urgent recovery work.
Decision challenge
The item-level and portfolio-level decisions are coupled. Giving one item more safety stock consumes budget that could protect another, so the problem is not only how much stock to hold—it is also which inventory model should govern each item.
Why the approach is fast
- Precompute, then select. Each candidate model first produces an item-level safety-stock quantity and a historical performance score. The portfolio optimizer then chooses among these fixed alternatives instead of repeatedly simulating the entire inventory system.
- Use a linear binary formulation. The selection and budget constraints form a binary integer linear program with special ordered set constraints. This lets mature solvers use specialized branching and cutting methods.
- Avoid nonlinear search at scale. The benchmark nonlinear approaches repeatedly approximate gradients and evaluate objectives and constraints; their runtime grows rapidly and convergence can become difficult as the item count reaches the thousands.
Why decision quality improves
The method does not claim to improve demand-forecast accuracy. It improves the safety-stock decision by allowing different items to use different models and by evaluating whether historical protection has enough margin to remain useful when demand changes.
- Model diversity: the public evaluation combines five candidates, including an MRP-based rule, normal-demand formulas, and empirical distribution-free formulas.
- Margin-aware scoring: models that merely tie at a 100% historical win rate are separated by how much additional demand they could absorb before a shortage.
- Diminishing returns: a root-margin score rewards useful protection but discounts excessive stock, leaving budget available for items where it reduces shortage risk more effectively.
Published practical results
The paper reports a three-month retrospective evaluation using an industrial inventory dataset, five candidate models, and a 12-month historical window. Compared with the tailored MRP-based benchmark, the root-margin version of SSMSO achieved:
- 50–60% fewer shortage-affected items per month.
- 51–64% lower adjusted shortage rates across the three monthly evaluation periods.
- Budget-feasible portfolio selection, keeping the combined safety-stock recommendation within the shared cap while several single-model alternatives exceeded it.
- About 800× faster optimization at 1,000 items: 1.72 seconds versus 23 minutes for one nonlinear benchmark.
Disclosure note
This page is an independent summary of results reported in the IEEE CASE 2025 paper and is not an official statement by my employer. Only information disclosed in the published paper is summarized; confidential data, client information, internal system details, proprietary implementation, and non-public operational results are omitted.
Reported improvements are from the paper’s retrospective evaluation and should not be interpreted as verified production-deployment outcomes. The adjusted shortage rate is a proxy for inventory performance.
This page uses independently written summaries and does not reproduce IEEE text, tables, figures, or the version-of-record article.
Public artifact
The method and evaluation appear in “Safety Stock Model Selection Optimization for Budget-Constrained Multi-Item Inventory Management: A Scalable Framework,” presented at IEEE CASE 2025. DOI ↗